git:20260715.61c1d82 to git:20260716.a78ef59

- <!-- rootloom:managed-start version=1 fingerprint=50ce962cc909f27dd01e scope=. -->

- # Project guidance for Rootloom

-

- ## Scope and sources of truth

-

- - Project purpose: Rootloom Personal Core keeps the part an individual developer uses every day (from `README.md`).

- - Detected sources of truth: `CONTRIBUTING.md`, `CONTRIBUTING.zh-CN.md`, `README.md`, `README.zh-CN.md`, `docs/architecture.md`, `docs/architecture.zh-CN.md`, `Makefile`, `.github/workflows/ci.yml`, `.github/workflows/codex-compatibility.yml`.

-

- ## Repository map

-

- - `docs/` — canonical project documentation.

- - `scripts/` — project automation.

- - `tests/` — test suites.

-

- ## Canonical commands

-

- - `make validate` — detected from `Makefile target validate`.

- - `make test` — detected from `Makefile target test`.

- - `make check` — detected from `Makefile target check`.

-

- ## Verification contract

-

- - Run the smallest detected command set that proves the changed behavior, then expand with blast radius.

- - If a required command cannot run, report the exact gap and do not convert it into a passing claim.

- - Keep generated guidance factual: project-only invariants belong below the managed block and must cite real paths.

- <!-- rootloom:managed-end -->

-

- <!-- Add durable project-only rules below this line. The seeder preserves content outside the managed block. -->

-

- ## Project-specific invariants

- - `plugins/rootloom/` is the only installable plugin source; `.agents/plugins/marketplace.json` must continue to point to that exact directory.

- - `plugins/rootloom/skills/seed-project-guidance/scripts/seed_project_guidance.py` owns automatic evidence collection and must remain deterministic, bounded, standard-library-only, network-free, repository-contained, lock-serialized, and snapshot-preserving.

- - `plugins/rootloom/skills/refine-project-guidance/SKILL.md` owns semantic project-guidance judgment; do not move model-dependent inference into the `SessionStart` Hook or deterministic scanner.

- - `plugins/rootloom/skills/record-engineering-decision/SKILL.md` owns portable decision memory; create records only for accepted durable choices and keep `AGENTS.md` references concise rather than duplicating rationale.

- - `plugins/rootloom/assets/system/AGENTS.md` owns persistent Standard plus task-scoped Single action/Full authorization semantics; static Rules must not recreate per-command prompts. See `docs/decisions/2026-07-14-tiered-authorization-modes.md`.

- - `plugins/rootloom/skills/setup-rootloom/scripts/setup_rootloom.py` owns Personal Core Codex-home writes and capability-to-artifact mapping; it must remain plan-first, ordinary-lock-serialized, mode-preserving on rollback, conflict-refusing, atomic per file, backed up, and hash-aware without claiming whole-transaction crash recovery.

- - `plugins/rootloom/hooks/run_component_hook.py` owns lifecycle-Hook enablement. Absent, malformed, or symlinked policy must fail closed; setup is the only supported way to enable a component.

- - `plugins/rootloom/skills/engineering-change/scripts/finalize_change.py` owns the lightweight personal review bundle; changes to changed-path capture, dangerous-deletion confirmation, command execution, output bounds, or summary format require focused regression coverage.

- - `plugins/rootloom/lib/rootloom_paths.py` owns separate secret-material privacy and security-domain risk classification, and `plugins/rootloom/lib/rootloom_memory.py` owns strict Engineering Memory reads, schema, identity, relevance, and lifecycle semantics. Capture must use only the material classifier for redaction; Analyzer uses the domain classifier for risk. Consumers must not grow divergent copies. See `docs/decisions/2026-07-15-sensitive-material-and-capture-bounds.md`.

- - `plugins/rootloom/skills/engineering-change/scripts/begin_review.py`, `seal_contract.py`, and `runner/{baseline,change_contract,review_run,state,strict_json}.py` own the opt-in strict-review draft/seal lifecycle, stable repository capture, sensitive-change quarantine, repository-base binding, strict evidence decoding, and machine scope/verification contracts. Keep them lightweight, local, bounded, standard-library-only, unable to authorize commands by file content alone, and absent from ordinary installation/task paths. See `docs/decisions/2026-07-14-personal-intelligence-contract.md`.

- - `plugins/rootloom/skills/engineering-change/scripts/runner/evidence_paths.py` owns lexical evidence/output paths and repository-storage containment. Both the worktree and resolved Git common directory are repository-owned, including for linked worktrees.

- - `plugins/rootloom/skills/engineering-change/scripts/runner/intelligence.py` owns advisory risk floors and verification planning; it must remain local, explainable, bounded, standard-library-only, subordinate to semantic evidence, and unable to auto-execute suggestions. See `docs/decisions/2026-07-14-personal-intelligence-contract.md`.

- - Strict Review summary revision 4, redaction caps, Git time/path budgets, targeted sensitive discovery, and exact seal recovery must remain evidence-honest and fail closed. See `docs/decisions/2026-07-15-evidence-honest-strict-review.md`.

- - `plugins/rootloom/skills/project-memory/scripts/project_memory.py` owns optional `.project-memory/` formats; updates must remain explicit, reviewable, standard-library-only, and subordinate to current repository evidence. Keep its advisory/relevance boundary aligned with `docs/decisions/2026-07-14-personal-intelligence-contract.md`.

- - The complete pre-split Assurance 1.2.19 product is retained on `codex/enterprise-assurance`; do not reintroduce its approval, audit, strict-runner, or recovery machinery into the `main` plugin by default.

- - Changes to installation, public behavior, safety gates, or user configuration must update both `README.md` and `README.zh-CN.md`.

- - `scripts/validate_repo.py` is the repository contract gate. Extend it when adding public assets, workflows, or manifest fields instead of relying only on prose review.